17

There's been a lot of talk about how much water AI uses. I keep reading that they use it for cooling, but after they do that where's the water going? From my understanding, liquid cooled PCs are fully enclosed not requiring top-ups unless there are leaks. Even if they do flow water in and back out, that wouldn't consume any water.

nbro
  • 42,615
  • 12
  • 119
  • 217
Ora G. Walters
  • 171
  • 1
  • 5

3 Answers3

21

The heat generated by large (AI) data centers is orders of magnitude greater than that of a liquid-cooled PC, which requires systems that can handle massive volumes of heat, such as water evaporative cooler into open air or discharged as wastewater. Efforts are underway to reduce water usage and improve efficiency, but evaporative cooling still remains a major component.

Evaporative cooling differs from other air conditioning systems, which use vapor-compression or absorption refrigeration cycles. Evaporative cooling exploits the fact that water will absorb a relatively large amount of heat in order to evaporate (that is, it has a large enthalpy of vaporization). The temperature of dry air can be dropped significantly through the phase transition of liquid water to water vapor (evaporation). This can cool air using much less energy than refrigeration. In extremely dry climates, evaporative cooling of air has the added benefit of conditioning the air with more moisture for the comfort of building occupants.

cinch
  • 11,000
  • 3
  • 8
  • 17
8

On top of the already mentioned cooling, following factors also play a role:

  • The production of semiconductors in general. According to World Economics Forum [1]

    An average chip manufacturing facility today can use 10 million gallons of ultrapure water per day—as much water as is used by 33,000 US households every day.

    Factoring in the overall need for a lot of technology, not just the AI chips itself, but a huge amount of other chips, memory, cables, for the main components, but also supporting components (such as highly technical systems for maintaining data center temperature, access mechanisms, network infrastructure, etc). In general, building and operating a data center itself is highly resource-intense.

  • The massive power consumption running large data centers has an indirect, yet high impact on water consumption (+ other environmental aspects). For similar reasons, factoring in everything that is necessary in order to generate electricity, a MWh of electricity can cost up to 1 million liters of water, depending on the method [2].

Generally, it's about the whole value chain of creating an AI service, from the physical technical components, to operation of IT or training the model. For example, training ChatGPT-3 apparently lead to consumption of 5.4 million litres of water [3], 700k litres of which were used for cooling the data centers.

[1] https://www.weforum.org/stories/2024/07/the-water-challenge-for-semiconductor-manufacturing-and-big-tech-what-needs-to-be-done/

[2] https://visualizingenergy.org/what-methods-of-electricity-generation-use-the-most-water/

[3] https://arxiv.org/abs/2304.03271

nbro
  • 42,615
  • 12
  • 119
  • 217
kopaka
  • 181
  • 2
0

Any production of electricity to plug in computers and other items requires large use of water. The mere fact an item takes an enourmous amount of ELECTRICITY means it is utilizing large amounts of water. Nuclear Power plants are on shorelines as water is used and dumped. A lot of the uranium 228 and uranium 226 in our WATER supplies that we DRINK is from power plants which are allowed to dump it.

Taylor
  • 1